VynFi/vynfi-group-audit-enterprise-2000

VynFi Group Audit — Enterprise 2000-entity (v5.33.2)

End-to-end synthetic group-audit bundle: 2 000 entities across North America, EMEA, and APAC, organised under a 4-level ownership graph, with full consolidated FS, intercompany eliminations, NCI rollforward, currency translation, and per-entity audit workpapers.

Refreshed on engine v5.33.2 (commit 7376c8fc) — picks up the three v5.33 fixes that the prior v5.10 dataset didn't have:

• #162: per-entity TB writer is now framework-aware (per-line account_type and category resolved via FrameworkAccounts::classify_account_type for the entity's framework).
• #164: consolidated FS aggregator inherits the same framework-aware classification via the new AggregatedAccount.account_type field.
• #163: balance/opening_balances.json now persists on every shard.

Plus every SOTA generator that landed between v5.10 and v5.33.2 (SOTA-1…12 + 8.1/11.1, all the v5.30 Tier-A/B work, v5.31 streaming aggregate, v5.32 calibration scaffolding).

All data is synthetic. No client names, no real corpus content.

Ownership structure

ACME_HQ (US, flagship)
├── ACME_EUR (DE, flagship, 100%)
│   ├── 25× ACME_EU_SIG (EU, 55-100%)
│   └── 350× ACME_EU (EU, 55-100%)
├── ACME_UK (GB, significant, 100%)
├── ACME_JP (JP, significant, 85%)
├── 50× ACME_NA_SIG (NA, 65-100%)
├── 700× ACME_NA (NA, 55-100%)
├── 200× ACME_AS (APAC, 51-100%)
├── 471× ACME_SMALL (BR/AR/ZA/AU, 50-100%)
└── 200× ACME_JV (BR/AR/ZA, 20-49%, equity-method)

= 1 + 3 + 25 + 350 + 50 + 700 + 200 + 471 + 200 = 2000 entities.

Period: 2024 Q1 (Jan-Mar), quarterly close. Presentation currency USD.

What's in the box

enterprise_2000_v533_2.tar.zst (1.24 GB, ~20 GB extracted) — full group archive:

enterprise_2000_v533_2/
├── manifest.json                       # group manifest with all 2000 entities
├── shard_summary.json                  # per-shard entity counts + IC JE counts
├── entities/<2000 dirs>/               # per-entity JEs, TBs, openings, FS, etc.
├── consolidated/                       # group-level
│   ├── consolidated_financial_statements.json
│   ├── consolidation_schedule.json
│   ├── nci_rollforward.json            # 200 equity-method + minority interests
│   ├── cta_rollforward.json            # FX translation worksheet
│   ├── equity_method_investments.json  # 200 equity-method entities
│   ├── translation_worksheet.json
│   ├── notes_to_consolidated_fs.json
│   ├── je_network.csv                  # 1,589,934 consolidated JE-line edges
│   └── je_network.parquet
└── ic_eliminations/
    └── ic_matching_coverage.json

Key statistics

• 2 000 entities, 2 154 intercompany relationships, 4 shards
• 1 800 entities summed (Parent + Full method); 200 deferred for equity-method overlay
• 9 118 intercompany journal entries
• 115 309 trial-balance lines aggregated, 604 unique GL accounts
• 1 589 934 consolidated JE-line edges in je_network.parquet
• 2 000 / 2 000 balance/opening_balances.json files persisted (was 0 / 2 000 in the v5.10 dataset — closes #163)
• Consolidated A vs L+E+NCI gap: 3.09 % (single-period; equity-method deferred entities contribute by design)
• IC matching coverage: 0.9159 (gap from IC relationships involving the 200 equity-method-deferred entities, expected by design)
• Engine: v5.33.2 commit 7376c8fc
• Generation walltime: ~1 min 44 s on a 30-core / 222 GB Xeon 8358 (down from the v5.10 multi-hour run via the v5.31 streaming- aggregate refactor)

Engine configuration

The full config is committed at docs/baselines/2026-05-28-v5.33.2-enterprise-2000/enterprise_2000.yaml in the engine repo. Pillar Two, CbC report, transfer-pricing master file, KAMs, group opinion all enabled.

Caveats

• Single period. This is one quarter (Q1 2024). For a multi-period chain with prior-year carry-forward see the companion VynFi/vynfi-group-audit-3yr-medium dataset (3 entities × 3 years).
• NCI in nci_rollforward.json, not BS line items. The v5.33.1 classifier's 3500-3599 NCI carve-out is a US-style convention; this group uses generated codes that don't hit that range. The controlling-vs-minority split lives in the rollforward file.
• 3.09 % consolidated BS residual comes from the 200 equity-method entities being held out of the pre-elim sum but still contributing to the equity-method-investment line on the parent's BS. The NCI / equity-method overlay applied by apply_nci_and_equity_method partially closes this but not all the way to 0. The per-entity TBs are individually balanced (is_balanced: true).
• Sign convention between closing_balance (debit-normal in TBs) and opening_balances::balances (natural-side) — same as documented on the 3-year medium dataset card.

How to use

pip install huggingface_hub
hf download VynFi/vynfi-group-audit-enterprise-2000 --repo-type dataset --local-dir ./e2k
tar -I zstd -xf ./e2k/enterprise_2000_v533_2.tar.zst

import pandas as pd, pyarrow.parquet as pq
edges = pq.read_table("enterprise_2000_v533_2/consolidated/je_network.parquet").to_pandas()
print(f"{len(edges):,} edges, {edges.is_fraud.mean():.4%} fraud-flagged")

Companion datasets

• VynFi/vynfi-group-audit-3yr-medium — multi-period chain demo (3 entities × 3 years, v5.33.2)
• VynFi/vynfi-je-network-2k — 68.5 M-edge JE-line graph in parquet

License

Apache-2.0.